Baker - Closing the Loop
Using educational data mining to design tutor lessons that students don’t choose to game: “Closing the loop”
Summary Table
| PIs | Ryan Baker |
| Other Contributers | |
| Study Start Date | Spring, 2010 |
| Study End Date | |
| LearnLab Site | TBD |
| LearnLab Course | Algebra |
| Number of Students | TBD |
| Total Participant Hours | TBD |
| DataShop | TBD |
Abstract
This 12 month CMDM project proposes to “close the loop” on a data mining analysis previously conducted within the PSLC (Baker_Choices_in_LE_Space), showing that the previous analysis makes a contribution to improving student learning in in-vivo settings. In that previous study, a model of the differences between different tutor lessons (the Cognitive Tutor Lesson Variation Space, or the CTLVS1 -- full details on this model are given on the page Baker_Choices_in_LE_Space) was created, and used to study why some tutor lessons are gamed more than others in the Algebra tutor. The best model based on the CTLVS1 (developed via a combination of PCA and correlation mining) predicted over half of the variance in gaming, almost 6 times better than any previous model attempting to explain gaming through specific student individual differences.
In this study, we will choose a lesson from the Algebra tutor that is highly gamed, and modify it in accordance with the findings of that previous work, such that the modified lesson is predicted to lead to significantly less gaming.
Background & Significance
In recent years, there has been considerable interest in how students choose to interact with learning environments. At any given learning event, a student may choose from a variety of learning-oriented "deep" paths, including attempting to construct knowledge to solve a problem on one’s own (Brown and vanLehn, 1980), self-explaining (Chi et al, 1989; Siegler, 2002), and seeking help and thinking about it carefully (Aleven et al, 2003). Alternatively, the student may choose from a variety of non-learning oriented "shallow" strategies, such as Help Abuse (Aleven & Koedinger, 2001), and Systematic Guessing (Baker et al, 2004). This pair of strategies is referred to as Gaming the system (Baker et al, 2004). Gaming the system is an active and strategic type of shallow strategy known to occur in many types of learning environments (cf. Baker et al, 2004; Cheng and Vassileva, 2005; Rodrigo et al, 2007), including the Cognitive Tutors used in LearnLab courses (Baker et al, 2004).
Recent work has indicated that a variety of aspects of cognitive tutor lessons are predictive of greater quantities of gaming (Baker et al, 2009). In particular, gaming is predicted by (boldface indicates a particularly strong relationship -- r2>0.15)
- Lack of text in problem statements not directly related to the problem-solving task, generally there to increase interest
- Not immediately apparent what icons in toolbar mean
- The lesson is not an equation-solver unit
- The location of the first problem step does not follow conventions (such as being the top-left cell of a worksheet) and is not directly indicated
- The same number being used for multiple constructs
- Reading hints does not positively influence performance on future opportunities to use skill
- Proportion of hints in each hint sequence that refer to abstract principles
- Hints do not give directional feedback such as “try a larger number”
- Hint requests that student perform some action
Historically, work to address gaming the system has either attempted to eliminate gaming behavior by making it more difficult to game (for instance, by putting delays between hints), or has attempted to detect gaming automatically and respond to it. The first approach appears to lead students to find new gaming strategies (Murray & VanLehn, 2005), and the second approach has led to systems that reduce gaming (Arroyo et al, 2007; Baker et al, 2006; Roll et al, 2007; Walonoski & Heffernan, 2006) and improve learning (Arroyo et al, 2007; Baker et al, 2006), but has typically been time-consuming and difficult to scale.
Within this project, we propose to use the findings from Baker et al (2009), which found tutor lesson features associated with gaming, as design principles for how to develop tutor lessons that students do not choose to game, in a completely unnoticeable way.
This study will also give us the opportunity to make a contribution to another area of research – recent work has suggested that interest-increasing text may reduce the transferability of knowledge. This may form a trade-off, where interest-increasing text both reduces gaming (improving learning) but reduces the transferability of the learning. We will investigate this question with conditions that both include interest-increasing text, and lack interest-increasing text, but are identical in all other ways.
Glossary
Computational Modeling and Data Mining
Hypotheses
- H1
- Re-designing a tutor lesson to eliminate features shown to be associated with gaming (see above), will result in lower gaming and better learning
- H2
- The tutor lesson features associated with gaming(see above) influence gaming by increasing the incidence of boredom and confusion
- H3
- Adding interest-increasing text to a tutor lesson (one such feature) will reduce gaming and improve learning, but potentially at the cost of lower transfer
Independent Variables
Three conditions will be compared:
- An unmodified version of a highly-gamed lesson in Cognitive Tutor Algebra
- A modified version of the same lesson
- Clearer communication of the flow of problem-solving through the interface (with a giant arrow)
- Fewer help messages that are wholly abstract in nature
- A second modified version of the same lesson
- Adding interest-increasing extraneous text to the scenario
- Clearer communication of the flow of problem-solving through the interface (with a giant arrow)
- Fewer help messages that are wholly abstract in nature
Dependent Variables
- Incidence of gaming behavior (measured via quantitative field observations -- cf. Baker et al, 2004)
- Incidence of boredom and confusion (measured via quantitative field observations -- cf. Rodrigo et al, 2007)
- Learning of domain skills and concepts (measured pre-post)
- Transfer and preparation for future learning (measured at post-test)
